1. Introduction
With the rise of smart video surveillance, a large number of people-tracking applications, autonomous vehicles, facial detection, and fast and accurate object detection methods are in ever-growing demand [
1,
2]. The continuing development in the areas of computer vision (CV), image processing, and deep learning (DL) techniques have changed the way we think about various characteristics of daily life [
3]. The DL methodology has provided a reliable basis for image recognition with reliable accuracy [
4]. The prevalent image classification, the convolutional neural network (CNN), is a fascinating biological neural network that is composed of distinct layers, with the neurons of all the layers being strongly associated with the neuron in the subsequent layer [
5]. The benefits of using a CNN are that they allow an independence between the feature extraction of preceding knowledge and a minimal design effort. CNN has made greater accomplishments in image recognition and classification [
6]. The popularity and accuracy of CNN for image classification have been optimized due to the largescale system for learning and image processing, higher-speed GPUs, and the huge availability of public datasets of an image [
7]. The idea of smart waste classification with waste and trash images has tremendous potential.
Owing to faster urbanization, currently, cities are facing significant challenges [
8,
9]. Amongst these challenges are those associated with the waste management system as the quantity of waste is directly proportional to the group of people living in urban regions. Waste management technology chiefly concerns the treatment and disposal of various kinds of waste [
10] and therefore protects animals, human beings, and the surrounding areas [
11]. An appropriate waste management technique could save money and result in less environmental pollution and improved air quality. Advanced regions of the world have been simultaneously implementing and discovering effective technologies for large-scale construction and effective waste management [
12,
13]. It is not possible to handle such a large quantity of waste in the forthcoming five years of the prevailing situation. Therefore, it is better to take each essential action needed for the effectual waste management [
14] and to adopt the best practices and techniques to efficiently treat waste and obtain a healthier environment.
Gokulnath et al. [
15] designed a new genetic approach and support vector machine (GASVM) to predict complex patterns, with the utilization of this approach offering improved results. Salama et al. in [
16] developed a learning neural network with an ant colony optimization (LNNAC) model for the prediction of complex patterns in automated recognition processes. Next, Yadav et al. [
17] have presented a new artificial neural network with a particle swarm optimization (PSO-ANN) model for resolving Troesch’s problem. The authors in [
18] introduced a new urban waste management technique which makes use of a cuckoo search-optimized long short-term recurrent neural network (CLST-RNN).
Malik et al. [
19] have presented a network by which to classify litter into types identified from benchmark techniques. The network they utilized to classify litter was EfficientNet-B0. Their study presents an EfficientNet-B0-based approach to the tuning of detailed images to specific demographic regions and from there to effective classification. This kind of tuning method on transfer learning (TL) offers a modified approach to classification, one which is extremely optimized for specific regions. Alsubaei et al. [
20] established an approach which mostly concentrates on the identification and categorization of lesser garbage waste objects in order to support an intelligent waste management system. In order to recognize an object, an improved RefineDet (IRD) technique with hyperparameter tuning process was employed. Secondly, a functional link neural network (FLNN) approach is executed to classify waste objects into several classes.
Verma et al. [
21] have presented a DL-based intelligent garbage recognition method. An objective purpose of their work was to manage garbage efficiently. In order achieve this, automation was developed utilizing two CNN techniques and images of solid waste that were taken by drones. Both of the CNN techniques were trained on the gathered image dataset at distinct rates of learning, optimization, and periods. Yang et al. [
22] have examined a new incremental learning structure, GarbageNet, in order to address the abovementioned challenges. Firstly, weakly-supervised transfer learning guarantees the ability of feature extraction. Secondly, to classify types of garbage, GarbageNet embeds them as anchors for reference and classifies the test instances by finding their nearest neighbors from the latent space. Thirdly, a considered collection of trained data were employed to suppress the negative outcomes of mislabeled data.
Kumar and Buelaevanzalina [
23] have provided a visual geometry group–neural network (VGG16-NN) technique that is dependent upon the procedure of attention to classified recyclable waste. Their attention module was established in order to model the important data from the feature map and provide increased detail. This technique automatically extracts classification features, namely organic, recyclable and non-recyclable waste. Kumar et al. [
24] have examined a new technique to waste segregation in order to achieve their effective recycling and disposal by employing a DL technique. The YOLOv3 technique was employed from the Darknet structure to train a self-made dataset.
In [
25], the authors employ a DL-based classification and CC system for realizing higher accuracy waste classification, beginning with garbage collection. To assist the subsequent waste disposal, the authors subdivide recyclable waste into glass, plastic, cardboard or paper, fabric, metal, and other recyclable wastes. DL-based A CNN is employed to realize the task of classifying garbage. Uganya et al. [
26] presented an automatic system for achieving an effectual and intelligent waste management scheme utilizing the IoT by forecasting the probability of waste items. The gas level, metal level, and wastage capacity were observed while utilizing IoT-based dustbins that could be located anywhere in city. Afterward, the authors presented techniques that had been tested by ML classifier approaches such as LR, linear regression, SVM, RF, and DT techniques. Though several models for waste classification are available in the literature, there is still a need to improve the performance of detection. At the same time, the trial-and-error hyperparameter tuning of the DL models is a tedious process. Therefore, metaheuristic optimization algorithms can be used for automated hyperparameter tuning.
This study develops a modified cuttlefish swarm optimization with machine learning-based solid waste management (MCSOML-SWM) in smart cities. In the MCSOML-SWM model, a single shot detector (SSD) model allows effectual recognition of objects. Then, a deep convolutional neural network (DCNN)-based MixNet model is applied to produce feature vectors and the hyperparameter tuning process is carried out by the MCSO algorithm. For accurate waste classification, the MCSOML-SWM technique applies a support vector machine (SVM) in this study. A comprehensive set of simulations were carried out to demonstrate the improved classification performance of the MCSOML-SWM model. In summary, the key contributions of the study are given as follows.
An intelligent MCSOML-SWM technique composed of an SSD object detector, a MixNet-based feature extraction, an MCSO-based parameter tuning, and an SVM classifier is presented. To the best of our knowledge, the MCSOML-SWM model has never been presented in the literature.
A novel MCSO algorithm is derived for hyperparameter tuning of the MixNet model.
Hyperparameter optimization of the MixNet model using MCSO algorithm using cross-validation helps to boost the predictive outcome of the MCSOML-SWM model for unseen data.
3. Performance Validation
The proposed model was simulated using Python 3.6.5 tool on PC i5-8600k, GeForce 1050Ti 4 GB, 16 GB RAM, 250 GB SSD, and 1 TB HDD. The parameter settings are given as follows: learning rate, 0.01; dropout, 0.5; batch size, 5; epoch count, 50; and activation, ReLU. In this study, the waste classifier results of the MCSOML-SWM method were tested using the TrashNet dataset [
31], which includes 2527 samples under six classes as represented in
Table 1. The class labels are glass, paper, cardboard, plastic, metal, and trash. The pictures were captured by placing the object on a white posterboard and using sunlight and/or room lighting. The pictures were resized to 512 × 384 pixels and the devices used were Apple iPhone 7 Plus, Apple iPhone 5S, and Apple iPhone SE.
A set of confusion matrices formed by the MCSOML-SWM technique under diverse epochs are represented in
Figure 1. The figure shows that the MCSOML-SWM system demonstrated improved waste classifier results under all epochs.
Table 2 and
Figure 2 highlight the waste classification outcomes of the MCSOML-SWM model on 200 epochs. The result shows that the MCSOML-SWM technique has provided improved outcomes under all classes. For example, in the glass class, the MCSOML-SWM method has offered outcomes for
,
,
,
, and
of 98.54%, 94.62%, 98.20%, 96.38%, and 93.01%, respectively. Additionally, for the paper class, the MCSOML-SWM approach provided outcomes for
,
,
,
, and
of 97.74%, 95.43%, 94.95%, 95.19%, and 90.82%, respectively. Furthermore, for the plastic class, the MCSOML-SWM method has provided outcomes for
,
,
,
, and
of 98.42%, 95.29%, 96.47%, 95.88%, and 92.08%, respectively.
Table 3 and
Figure 3 highlight the waste classification outcomes of the MCSOML-SWM method on 400 epochs. The results demonstrate that the MCSOML-SWM method has provided enhanced outcomes under all classes. For instance, in the glass class, the MCSOML-SWM model has offered outcomes for
,
,
,
, and
of 99.25%, 97.07%, 99.20%, 98.12%, and 96.32%, respectively. Furthermore, for the paper class, the MCSOML-SWM model has given outcomes for
,
,
,
, and
of 99.13%, 98.81%, 97.47%, 98.14%, and 96.34%, respectively. Moreover, in the plastic class, the MCSOML-SWM model has provided outcomes for
,
,
,
, and
of 99.37%, 97.94%, 98.76%, 98.35%, and 96.75%, respectively.
Table 4 and
Figure 4 highlight the waste classification outcomes of the MCSOML-SWM method on 600 epochs. The results show that the MCSOML-SWM approach has provided better outcomes under all classes. For example, in the glass class, the MCSOML-SWM technique has provided outcome for
,
,
,
, and
of 98.73%, 95.36%, 98.40%, 96.86%, and 93.90%, respectively. Furthermore, for the paper class, the MCSOML-SWM model has given outcomes for
,
,
,
, and
of 97.90%, 95.77%, 95.29%, 95.53%, and 91.44%, respectively. Furthermore, for the plastic class, the MCSOML-SWM model has given outcomes for the
,
,
,
, and
of 98.73%, 95.92%, 97.51%, 96.71%, and 93.63%, respectively.
Table 5 and
Figure 5 highlight the waste classification outcomes of the MCSOML-SWM system on 800 epochs. The results show that the MCSOML-SWM algorithm has provided better outcomes under all classes. For example, in the glass class, the MCSOML-SWM technique has offered outcomes for the
,
,
,
, and
of 98.85%, 95.38%, 99%, 97.16%, and 94.48%, respectively. Moreover, for the paper class, the MCSOML-SWM model has given outcomes for the
,
,
,
, and
of 97.78%, 96.38%, 94.11%, 95.23%, and 90.89%, respectively. Furthermore, for the plastic class, the MCSOML-SWM model has presented outcomes for the
,
,
,
, and
of 98.73%, 96.11%, 97.30%, 96.70%, and 93.61%, respectively.
Table 6 and
Figure 6 highlight the waste classification outcomes of the MCSOML-SWM algorithm on 1000 epochs. The results show that the MCSOML-SWM method has provided enhanced outcomes under all classes. For instance, in the glass class, the MCSOML-SWM model has presented outcomes for
,
,
,
, and
of 98.93%, 95.75%, 99%, 97.35%, and 94.84%, respectively. Furthermore, for the paper class, the MCSOML-SWM model has provided outcomes for
,
,
,
, and
of 97.98%, 95.48%, 95.96%, 95.72%, and 91.79%, respectively. Moreover, for the plastic class, the MCSOML-SWM model has given outcomes for
,
,
,
, and
of 98.77%, 96.30%, 97.30%, 96.80%, and 93.80%, respectively.
The training accuracy (TRA) and validation accuracy (VLA) accomplished by the MCSOML-SWM methodology on the test dataset are demonstrated in
Figure 7. The results demonstrate that the MCSOML-SWM algorithm has accomplished the highest values of TRA and VLA. Additionally, the VLA seemed to be improved over the TRA.
The training loss (TRL) and validation loss (VLL) attained by the MCSOML-SWM methodology on the test dataset are portrayed in
Figure 8. The results illustrate that the MCSOML-SWM method has accomplished minimum values of TRL and VLL. Particularly, the VLL shows lower values than the TRL.
A clear precision–recall assessment of the MCSOML-SWM approach on the test dataset is depicted in
Figure 9. The figure shows that the MCSOML-SWM methodology has resulted in improved values of precision–recall values under each class.
A brief ROC investigation of the MCSOML-SWM approach on the test dataset is depicted in
Figure 10. The results signify that the MCSOML-SWM methodology has shown its capability in classifying distinct classes on the test dataset.
A wide ranging comparison study of the MCSOML-SWM technique with other waste classifier models is portrayed in
Table 7 and
Figure 11 and
Figure 12. The comparison study demonstrated that the MCSOML-SWM model shows superior outcomes over other techniques. With respect to
, the MCSOML-SWM method has demonstrated higher
of 99.34% whereas the GA-SVM, LNNAC, PSO-ANN, CLST RNN, and CNN models have depicted lower values for
of 85.08%, 90.58%, 95.20%, 98.58%, and 98.09%, respectively. Meanwhile, with respect to
, the MCSOML-SWM model has established a high value for
of 97.97% while the GA-SVM, LNNAC, PSO-ANN, CLST RNN, and CNN methods have portrayed lower values for
of 88.18%, 88.68%, 92.95%, 97.08%, and 97.76%, respectively. Ultimately, with respect to
, the MCSOML-SWM model has demonstrated high outcomes for
of 97.41% while the GA-SVM, LNNAC, PSO-ANN, CLST RNN, and CNN models have shown lower values for
of 84.15%, 90.13%, 94.20%, 97.28%, and 96.15%, respectively. Finally, with respect to
, the MCSOML-SWM model has established a high value for
of 97.67% while the GA-SVM, LNNAC, PSO-ANN, CLST RNN, and CNN models have portrayed lower values for
of 88.85%, 91.45%, 93.93%, 97.35%, and 95.09%, respectively.
Therefore, the MCSOML-SWM model has surpassed all the other waste classification models in the smart city environment.